Detecting hate speech poses significant challenges, especially in low-resource languages like Arabic, which features intricate linguistic structures and a variety of dialects. This study offers a comparative analysis of hate speech detection within the Arabic-speaking population, concentrating on the South Iraq dialect as it appears on social media platforms like Telegram. The methodology adopted in this research integrates various natural language processing techniques for text representation, alongside classical classification algorithms, hyperparameter optimization, and ensemble machine learning methods. The classical algorithms employed include Logistic Regression (LR), Support Vector Machine (SVM), Naïve Bayes (NB), Decision Tree (DT), K-Nearest Neighbor (KNN), and MultiLayer Perceptron (MLP). Hyperparameter tuning is executed using the GridSearch method with K-Fold cross-validation, and five ensemble ML classifiers are utilized. Feature extraction is performed through both TF-IDF and Word2Vec, with results benchmarked against the advanced AraBERTv2 model. The research utilizes a dataset collected locally, consisting of 1,067 posts in the South Iraq dialect. Among these, 527 posts are categorized as ‘non-hate speech’ (class 0) and 526 as ‘hate speech’ (class 1). This dataset is a significant aspect of the study, as it sets a benchmark for detecting hate speech within the South Iraq dialect. Experimental findings reveal that the AraBERTv2 model outperforms others, achieving an impressive accuracy rate of 92%. Overall, this work not only contributes a valuable dataset for future research but also demonstrates the potential of advanced models like AraBERTv2 in addressing the challenges of hate speech detection in specific dialects.